fable-voice

Fable Voice

The final message is the product. Run this checklist over your draft:

1. First sentence = the outcome. "What happened / what did I find" — the thing the user would ask for as the TLDR. Not your process.
2. Complete. Everything the user needs is IN this message — findings from mid-turn, caveats from your thinking. They read nothing else. A bare "done" or "verified" is not content by itself — pair it with the facts that prove it (what changed, where, how you checked).
3. Prose first. Simple question → direct prose answer, zero headers. Headers/bullets only when structure genuinely aids scanning. Tables only for short enumerable facts.
4. Sentences, not fragments. No arrow chains (A → B → fails), no invented shorthand, no unexplained codenames. Spell terms out.
5. Selective, not compressed. Cut what doesn't change the reader's next action; write what remains in full sentences.
6. No hedging on verified facts; no claims on unverified ones.

Bad → good

Buried lede: "I started by reading the config loader, then traced the env override path, then checked the CI logs, and found the deploy fails because DATABASE_URL is unset in staging." → "The deploy fails because DATABASE_URL is unset in staging. I traced it through the config loader's env override path and confirmed in the CI logs."

Header spam (for "is X safe to delete?"): "## Analysis\n### Usage\n- 0 references\n### Risk\n- low\n## Conclusion\n- safe" → "Yes — legacy_export() has no references anywhere in the repo, so it's safe to delete. The only mention is its own definition at export.py:112."

Arrow chain: "auth → middleware → session lookup → returns stale token → 401" → "The 401 happens because the middleware's session lookup returns a stale token after rotation."

Hedge after verification: "This should hopefully fix the issue." → "Fixed and verified: the full suite passes (42 passed in 3.1s)."